If your team runs GPT-5.6 Terra or Luna through OpenAI's API, today's announcement is worth five minutes of your time. OpenAI just cut prices on two of its three GPT-5.6 models, which means the cost math behind workflows using Terra or Luna changed as of today, July 30, 2026. Before you assume your current setup is still the cheapest option, it's worth checking which tasks should move to a lighter model and which should stay put.
What Actually Changed on July 30
OpenAI announced the update in a post titled "Advancing the price-performance frontier with GPT-5.6." The headline changes:
- GPT-5.6 Luna costs 80% less, starting today.
- GPT-5.6 Terra costs 20% less, starting today.
- A new Fast mode replaces Priority Processing for GPT-5.6 Sol in the API, offering up to 2.5x faster responses than standard processing, at twice the price, with no change in intelligence. Requests already tagged priority move to Fast mode automatically.
The lower prices also apply to how usage is counted against paid subscriptions in Codex and ChatGPT Work. Subscription prices and quota budgets aren't changing, but Terra and Luna tasks now draw down fewer credits against those quotas, so the same plan effectively stretches further. OpenAI says the new pricing begins rolling out on AWS later today as well.
Three Models, Now Cheaper to Match to the Job
GPT-5.6 comes in three tiers, and OpenAI is explicit that the point of the lineup is matching model intelligence to the task, not defaulting to the biggest model out of habit:
- Sol is the flagship, built for uncertainty and planning.
- Terra is the balanced, everyday-work model.
- Luna is the fastest and most affordable, aimed at high-volume work, tool use, and multi-step workflows.
Check out the new standard API pricing for short-context requests, per million tokens, per OpenAI's pricing page.
OpenAI also lists separate, higher rates for long-context requests. Batch and Flex processing have different rates as well. Use the pricing page for the exact rate that applies to your workload.
Luna, in particular, is now cheap enough to change the calculus for tasks that were previously routed to Terra or Sol mostly out of caution.
Which Workloads to Re-Evaluate First
OpenAI's own guidance is to weigh stakes, cost of error, urgency, and scale when picking a model, then let cheaper tiers take on more of the routine load. A practical way to apply that this week:
1. Move high-volume, well-specified work to Luna. Background agent automations, routine implementation, large-scale document analysis, and customer-interaction classification are exactly the categories OpenAI points to. At $0.20 input and $1.20 output per million tokens, Luna is now cheap enough to run at scale without the bill becoming the bottleneck.
2. Consider Terra for everyday Q&A and scoped tasks. OpenAI cites workspace Q&A and scoped tasks as Terra use cases through a Notion customer example. For many teams, that makes Terra a practical candidate for internal Q&A, workspace search, and clearly bounded requests, now 20% cheaper.
3. Reserve Sol for uncertainty and high stakes. Planning, ambiguous requirements, and anything where a wrong answer is expensive may still belong on the flagship model. OpenAI's own suggested pattern is to use Sol to resolve uncertainty and define a plan, then hand the well-specified execution to Luna, including writing and running tests and evaluating results.
4. Turn on Fast mode selectively. It's a 2x price premium for up to 2.5x the speed, with the same intelligence. That trade only makes sense when latency is the actual constraint, such as a live customer-facing interaction, not a background job that can finish overnight
Coursera next step: If this post has you thinking about AI spend as a leadership decision, AI Leadership & Strategic Implementation can help you connect model choices to rollout plans, governance, and business value. It is a practical next step for leaders who want AI adoption to be intentional, not just cheaper.*
Don't Switch Blindly, Run the Eval First
None of this is a reason to re-route every workload to Luna today. A cheaper model that produces worse output, needs more retries, or requires a human to redo the work isn't actually cheaper once you count total task cost. Before shifting production traffic, run a small evaluation on a representative sample of your actual prompts and compare output quality, latency, and end-to-end cost per completed task, not just the sticker price per token. Model behavior can shift even within the same family, so the safest move is to test before you commit.
Key Takeaway
GPT-5.6's price cuts widen the economic gap between the model tiers, making model selection a bigger factor in AI spending than it was yesterday. Business leaders don't need to overhaul their AI stack overnight, but they do need to check whether their current default model still makes sense.
Action checklist:
- Inventory current GPT-5.6 workloads and note which model each one runs on.
- Flag high-volume, well-specified tasks (batch processing, classification, routine implementation) as Luna candidates.
- Consider Terra for everyday Q&A and scoped internal tasks, instead of defaulting to Sol, unless stakes are high.
- Run a small quality and cost eval before moving any production workload to a cheaper tier.
- Enable Fast mode only for latency-sensitive, customer-facing requests.
- Monitor Codex and ChatGPT Work usage to see how the lower Terra and Luna credit consumption affects your available quota.
Sources
- Advancing the price-performance frontier with GPT-5.6, OpenAI
- GPT-5.6, OpenAI
- Pricing, OpenAI API Docs